Dynamic voiceprint authentication method and apparatus, computer device, and storage medium
By collecting and fusing voiceprint features in real time, and combining them with identity authentication and security verification, the problem of voiceprint drift in static voiceprint authentication is solved, enabling efficient and secure updates of dynamic voiceprint authentication, thus improving user experience and system security.
Patent Information
- Application Number
- PCT/CN2025/135632
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-06
- Filing Date
- 2025-11-18
- Publication Date
- 2026-11-12
AI Technical Summary
Existing static voiceprint authentication technology suffers from a decrease in matching rate due to voiceprint drift, requiring users to update frequently, increasing operational burden, and is vulnerable to attacks, affecting financial and medical security.
By collecting voice data in real time, extracting voiceprint features and fusing them with historical features, and employing a dual mechanism of identity authentication and voice security verification, dynamic voiceprint features are generated for authentication.
This enables seamless voiceprint updates, improves matching rates, reduces the number of re-entry attempts, enhances security, and ensures the operational efficiency of financial and medical services as well as user privacy.
Smart Images

Figure CN2025135632_12112026_PF_FP_ABST
Abstract
Description
Dynamic voiceprint authentication methods, devices, computer equipment and storage media
[0001] This application claims priority to Chinese Patent Application No. 202510581526.1, filed on May 6, 2025, entitled "Dynamic Voiceprint Authentication Method, Apparatus, Computer Equipment and Storage Medium", the entire contents of which are incorporated herein by reference. Technical Field
[0002] This application relates to the fields of artificial intelligence, information security, financial technology and digital healthcare, and in particular to dynamic voiceprint authentication methods, devices, computer equipment and storage media. Background Technology
[0003] Voiceprint authentication is a biometric authentication technology that verifies a speaker's identity by extracting, analyzing, and comparing voiceprint feature parameters based on the speaker's unique characteristics in their voice signal. As voiceprint authentication technology matures, biometric authentication based on voiceprint features is beginning to be applied and promoted in the field of identity authentication. This technology eliminates the need for direct physical contact between the user and the device; identity verification is completed through voice capture, offering significant advantages in convenience and security, and is particularly widely used in high-security fields such as finance and healthcare.
[0004] In the financial sector, voiceprint authentication technology has been deeply integrated into various online and offline service scenarios. For example, in mobile banking apps, key operations such as user login, fund transfer, and service activation can be authenticated via voiceprint, eliminating the need for cumbersome password input or SMS verification. Transactions can also be triggered by voice; for instance, a user can input "transfer 1000 yuan to Zhang San" and complete the transfer without manually clicking on interactive components, greatly improving user convenience. Wealth management platforms and online lending platforms can also use voiceprint authentication to identify users and effectively prevent unauthorized identity theft for financial transactions.
[0005] In the medical field, voiceprint authentication technology also demonstrates enormous application potential. In medical services and smart healthcare systems, patients can register for appointments via voice. This method is particularly convenient for patients with facial, eye, or hand injuries, as voiceprint authentication technology can accurately identify users, automatically linking their personal information and medical records, simplifying the registration process and reducing human input errors. When conducting online consultations or accessing test reports, voiceprint authentication ensures that only the individual can access relevant medical information, preventing the leakage of patient privacy.
[0006] However, current voiceprint authentication is mostly based on static voiceprint features, requiring users to pre-record fixed voice samples to form a static voiceprint database. Subsequent verification relies solely on the original voiceprint features. The inventors realized that over time, a user's voiceprint changes with age and health conditions, such as vocal cord swelling or inflammation caused by a cold. These changes significantly reduce the matching rate of the static voiceprint database. If a match fails, the user needs to re-record their voiceprint, increasing the user's workload, reducing user experience, and impacting the operational efficiency of financial and medical institutions. Furthermore, failing to update voiceprint information for extended periods makes it vulnerable to malicious actors using recording and AI synthesis techniques to bypass voiceprint verification. In the financial sector, this could lead to theft; in the medical field, it could result in patient privacy breaches, medical data tampering, disrupting normal medical order and endangering patients' lives and health, causing irreparable losses to users and institutions. Summary of the Invention
[0007] This application provides a dynamic voiceprint authentication method, apparatus, computer equipment, and storage medium to solve the technical problems of traditional voiceprint authentication based on static voiceprint features, which are difficult to overcome voiceprint drift and ensure security.
[0008] Firstly, it provides a dynamic voiceprint authentication method, including:
[0009] Real-time acquisition of input voice and triggering an authentication command when input voice is acquired;
[0010] Preprocess the input speech and extract speaker features;
[0011] The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features;
[0012] The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated.
[0013] Perform security verification on the input voice and generate the security verification result;
[0014] The identity authentication result is output based on the security verification result and the preliminary authentication result.
[0015] Secondly, a dynamic voiceprint authentication device is provided, including:
[0016] The acquisition module is used to acquire input voice in real time and trigger an authentication command when input voice is acquired;
[0017] The extraction module is used to preprocess the input speech and extract speaker features;
[0018] The fusion module is used to fuse the extracted voiceprint features with historical voiceprint features to generate the target voiceprint features;
[0019] The matching module is used to perform identity authentication matching on the target voiceprint features and generate preliminary authentication results;
[0020] The verification module is used to perform security verification on the input voice and generate the security verification result;
[0021] The authentication module is used to output the identity authentication result based on the security verification result and the preliminary authentication result.
[0022] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:
[0023] Real-time acquisition of input voice and triggering an authentication command when input voice is acquired;
[0024] Preprocess the input speech and extract speaker features;
[0025] The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features;
[0026] The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated.
[0027] Perform security verification on the input voice and generate the security verification result;
[0028] The identity authentication result is output based on the security verification result and the preliminary authentication result.
[0029] Fourthly, a non-volatile computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, performs the following steps:
[0030] Real-time acquisition of input voice and triggering an authentication command when input voice is acquired;
[0031] Preprocess the input speech and extract speaker features;
[0032] The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features;
[0033] The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated.
[0034] Perform security verification on the input voice and generate the security verification result;
[0035] The identity authentication result is output based on the security verification result and the preliminary authentication result.
[0036] The advantages of this application compared to existing technologies are as follows: This application collects voice and extracts voiceprint features in real time, fuses the extracted voiceprint features with historical features, and then uses a dual mechanism of identity authentication matching and voice security verification to output the authentication result. This allows for automatic voiceprint authentication and updating without the user's awareness, eliminating the need for active user voiceprint updates and significantly improving the user experience. Furthermore, it enables dynamic updating of voiceprint features, allowing the voiceprint authentication method to automatically adapt to changes in voiceprint caused by factors such as age and health. Compared to traditional static voiceprint databases, this significantly improves the matching rate, reduces the number of re-entry attempts, and substantially optimizes the user experience. It also ensures the operational efficiency of financial and medical services. The dynamic update mechanism enhances the security of the voiceprint database, reduces the risk of attacks due to outdated voiceprint information, and effectively protects user funds and privacy. Moreover, it comprehensively resists attacks on voiceprint authentication, building a security defense for financial transactions and medical data management, maintaining normal transaction order and the medical environment, and enhancing user and institutional trust in voiceprint authentication technology.
[0037] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of this application more obvious and understandable, the following are preferred embodiments, which are described in detail below. Attached Figure Description
[0038] Figure 1 is a schematic diagram of an application environment for a dynamic voiceprint authentication method according to an embodiment of this application;
[0039] Figure 2 is a flowchart of a dynamic voiceprint authentication method according to an embodiment of this application;
[0040] Figure 3 is a flowchart illustrating a specific implementation of step S20 in Figure 2;
[0041] Figure 4 is a flowchart illustrating a specific implementation of step S30 in Figure 2;
[0042] Figure 5 is a flowchart illustrating a specific implementation of step S50 in Figure 2;
[0043] Figure 6 is a flowchart illustrating a specific implementation of step S51 in Figure 5;
[0044] Figure 7 is a structural schematic diagram of a dynamic voiceprint authentication device in one embodiment of this application;
[0045] Figure 8 is a structural schematic diagram of a computer device according to an embodiment of this application;
[0046] Figure 9 is another structural schematic diagram of a computer device according to one embodiment of this application. Detailed Implementation
[0047] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0048] It should be understood that, when used in this specification and the appended claims, the terms “comprising” and “including” indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0049] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0050] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0051] Please refer to Figures 1 and 2. Figure 1 is a schematic diagram of an application scenario for the dynamic voiceprint authentication method provided in this application embodiment. During the use of the client, the user performs voice verification or voice control operations; the client communicates with the server via the network; the server can receive the input voice through the client, process the input voice, authenticate the user's identity based on the input voice, and then return the authentication result to the client; if the authentication is successful, the user can perform subsequent business operations through the client; if the authentication fails, the user needs to perform secondary authentication. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a separate server or a server cluster consisting of multiple servers. The following detailed description of specific embodiments further illustrates this application.
[0052] Please refer to Figure 2, which is a schematic flowchart of the dynamic voiceprint authentication method provided in this embodiment of the application. The dynamic voiceprint authentication method includes the following steps:
[0053] S10: Real-time acquisition of input voice and triggering an authentication command when input voice is acquired.
[0054] In this embodiment, the presence of input voice can be detected in real time via the client. If so, the input voice is acquired, and the identity authentication process is initiated. By acquiring voice in real time through step S10, voiceprint authentication and updates can be automatically completed in the background without the user's awareness, eliminating the need for the user to actively perform voiceprint update operations and significantly improving the user experience.
[0055] The dynamic voiceprint authentication method of this embodiment is applicable to various types of application platforms, especially financial or medical application platforms with high security requirements, such as online banking systems, securities trading systems, payment and settlement systems, online payment platforms, insurance business systems, financial data platforms, smart medical systems, patient information management systems, electronic medical record systems, medical image archiving and communication systems, telemedicine systems, etc.
[0056] For example, when operating an online banking system, users can proactively perform voice verification under the guidance of the client. Voice verification serves the same purpose as password verification, facial recognition, and fingerprint verification: to determine whether the person inputting the voice is the user of the current account. Only if so can the user log in or perform actual business operations. Users can also control actual business operations on the client via voice according to their needs. For example, if they say the transaction instruction "transfer 1000 yuan to Zhang San," the server will authenticate the user's identity. If the authentication is successful, it confirms that the person inputting the voice is the user of the current account, and after confirming the user's current operation is secure, it will respond to the instruction "transfer 1000 yuan to Zhang San." For example, when operating an electronic medical record system, users can input voice for voiceprint authentication to ensure that the person requesting to view their personal medical records is the actual user, preventing others from stealing their medical information and disseminating their privacy. When a user says "Book an appointment for the internal medicine clinic at 9 a.m. tomorrow" in a smart healthcare system, the server will authenticate the user's identity. If the authentication is successful, it confirms that the user who inputs the voice is the current user of the account, and after confirming that the user's current operation is secure, it can respond to the "Book an appointment for the internal medicine clinic at 9 a.m. tomorrow" command.
[0057] S20: Preprocess the input speech and extract voiceprint features.
[0058] In practical applications, user-inputted voice is inevitably affected by factors such as environmental noise and device differences. For example, a user might use an online banking system for voiceprint authentication in a noisy street, where surrounding vehicle horns and loud conversations can mix in the voice signal. In digital healthcare scenarios, patients using smart healthcare systems in hospital lobbies to register via voice can also experience issues with voice quality due to hospital announcements and footsteps. Noise reduction processing of the input voice can effectively remove these interfering signals, improve voice purity, and make the voice clearer, providing a high-quality voice signal for subsequent voiceprint feature extraction and significantly improving the accuracy and stability of voiceprint feature extraction. In the high-security fields of finance and healthcare, this operation can reduce misjudgments caused by poor voice quality, improve authentication efficiency, and ensure smooth business operations.
[0059] Voiceprint feature extraction involves extracting unique information that identifies the speaker from preprocessed speech signals, such as fundamental frequency, formants, and spectral envelope. These features are unique and stable, and can be matched with previously stored voiceprint features to confirm the speaker's identity. In financial applications, voiceprint feature extraction is used to verify user identity and ensure fund security. For example, in insurance systems, when a user applies for a claim, voiceprint feature extraction is compared with stored voiceprint information to prevent unauthorized individuals from fraudulently claiming insurance money through audio playback or AI-synthesized speech. In securities trading systems, voiceprint feature extraction ensures that trading instructions are issued by the account holder, preventing account theft and illegal trading. Voiceprint feature extraction is equally crucial in medical applications. For instance, in patient information management systems, when patients query their personal health records via voice, voiceprint feature extraction ensures that only the account holder can access the relevant information, protecting patient privacy.
[0060] In some embodiments of this application, as shown in FIG3, a specific preprocessing and feature extraction scheme is provided. In S20, that is, the input speech is preprocessed and the voiceprint features are extracted, which specifically includes the following steps S21-S22.
[0061] S21: Perform spectral subtraction noise reduction and frame segmentation on the input speech to obtain several speech segments.
[0062] Spectral subtraction denoising is a technique that reduces noise based on the difference in the spectral density between speech and noise. It works by assuming that the speech and noise signals are independent and removing noise by subtracting the estimated noise power spectrum from the power spectrum of the noisy speech. In practical applications, it can effectively reduce the interference of environmental noise on speech signals.
[0063] In input speech, the speech signal is a non-stationary signal that varies over time, but can be approximated as a stationary signal within a short period. The framing processing in this embodiment divides the continuous speech signal into a series of short speech segments. In this embodiment, the input speech is divided into segments with a frame length of 25ms and a frame shift of 10ms, allowing for subsequent analysis and processing of each segment, which helps improve the accuracy and efficiency of voiceprint feature extraction. For example, in an insurance business system, when a user makes a voice claim, the speech processed by spectral subtraction noise reduction and framing can more accurately extract voiceprint features, preventing misidentification due to noise interference and ensuring the security and accuracy of the claims process. In an electronic medical record system, when a patient queries their medical records by voice, the processed speech ensures accurate voiceprint authentication, avoiding the risk of medical record information leakage.
[0064] S22: Sequentially extract the Mel frequency cepstral coefficients and linear prediction coding features of each frame of speech segment to obtain several initial voiceprint features.
[0065] Mel-frequency cepstral coefficients (MFCCs) are a set of feature parameters extracted based on the Mel-frequency scale. Mel frequency is a frequency scale established based on the characteristics of human hearing, and it has a non-linear relationship with actual frequency. MFCCs can well simulate the human ear's perception of sounds at different frequencies and are widely used in speech recognition and feature extraction, effectively characterizing features such as timbre and pitch of speech. Linear predictive coding (LPC) is a method that extracts features by linearly predicting speech signals. It predicts the current input speech based on historical speech features and represents the features of the speech signal by solving for the prediction coefficients. It can effectively extract the spectral envelope of the speech signal, reflecting important information such as formants.
[0066] In this embodiment, the voiceprint features extracted using Mel-frequency cepstral coefficients and linear predictive coding exhibit good stability under various environmental and speech conditions. Furthermore, both methods possess strong discriminative power, clearly distinguishing the voiceprint differences between different speakers. In complex financial transaction environments, whether in noisy public places or quiet indoor spaces, these two features can effectively differentiate user identities, ensuring transaction security. In medical environments, even if different patients' voices are quite similar, the voiceprint features based on Mel-frequency cepstral coefficients and linear predictive coding can accurately distinguish them, avoiding confusion regarding medical information.
[0067] S30: The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features.
[0068] In this embodiment, historical voiceprint features are those stored in the corresponding user's voiceprint database prior to this identity authentication. These historical voiceprint features are accumulated and stored through multiple past authentication processes and possess relative stability over a certain period. They record relatively fixed vocal characteristics of the user, such as pronunciation habits and vocal tract structure. This stable feature information provides a reliable basis for voiceprint authentication. Even if the user's voiceprint changes, historical voiceprint features can still serve as an important comparison basis to help the system identify the user's identity. However, a user's voiceprint can change with age, health conditions, environmental factors, or device differences. Step S30 integrates newly extracted voiceprint features with historical voiceprint features, enabling the system to dynamically capture these changes. For example, in financial applications, a user's voiceprint may undergo subtle changes as they age. If authentication relies solely on the initially recorded historical voiceprint features, matching may fail. By incorporating new voiceprint changes into the system and continuously updating the voiceprint model, the system can adapt to the dynamic evolution of voiceprints. This ensures accurate user identification at different times, avoids authentication barriers caused by voiceprint changes, and improves the continuity and convenience of users' access to financial services. In medical applications, when a patient's voiceprint changes due to illness, such as vocal cord swelling or inflammation caused by a cold, the fused target voiceprint features can still ensure successful identity authentication in the smart healthcare system, allowing for normal operations such as registration, consultation, and medical record retrieval.
[0069] Newly extracted voiceprint features reflect the user's current voice state, while historical voiceprint features record the user's past voice information. Fusion of these two features integrates multi-dimensional voiceprint data, overcoming the limitations of single features. For example, in complex environmental noise, newly extracted voiceprint features may contain errors due to interference, while historical voiceprint features provide a stable reference. Through feature fusion, the system can optimize these data, mutually complementing and verifying each other, thereby more accurately characterizing the user's voiceprint features and improving the accuracy of identity authentication. Furthermore, the fusion of extracted and historical voiceprint features constructs a dynamically updated voiceprint model, enabling the voiceprint authentication system to have self-optimization capabilities. As the number of times users use the system increases and new features are continuously integrated, the voiceprint model becomes increasingly closer to the user's actual voiceprint, continuously improving authentication performance. Simultaneously, it reduces the frequency of users needing to manually re-enter their voiceprints due to changes, greatly enhancing the user experience. In the financial and medical fields, this efficient, accurate, and user-friendly authentication method helps improve industry service quality and user trust, promoting efficient business operations and healthy development.
[0070] In some embodiments of this application, as shown in FIG4, a specific feature fusion scheme is provided. In S30, the extracted voiceprint features are fused with historical voiceprint features to generate target voiceprint features, specifically including the following steps S31-S37.
[0071] S31: Obtain historical voiceprint features from the corresponding user's voiceprint database.
[0072] S32: Combine the initial voiceprint features and historical voiceprint features into a feature set.
[0073] For steps S31-S32, in this embodiment, all voiceprint features in the feature set are represented by feature vector v. t It means that v t ∈R d In this context, dimension d = 40. In voiceprint authentication technology, speech signals contain rich acoustic information, and voiceprint features are key elements extracted from speech signals that characterize the speaker's identity, such as Mel-frequency cepstral coefficients and linear predictive coding features. Representing these voiceprint features using feature vectors essentially transforms complex acoustic information into a mathematical structure that computers can efficiently process. Each feature vector is a sequence of multiple values, corresponding to the quantization results of the voiceprint feature across different dimensions. For example, in Mel-frequency cepstral coefficients, some dimensions may reflect the low-frequency components of speech, while others reflect the high-frequency components; in linear predictive coding features, some dimensions may be related to the formant characteristics of speech. This vector representation presents voiceprint features in a standardized and structured form, facilitating computer storage, computation, and comparison, laying the foundation for subsequent voiceprint authentication processes. Dimension d = 40 means that each feature vector consists of 40 values, and these 40 dimensions describe the voiceprint features from different perspectives. In the actual voiceprint authentication process, different dimensions correspond to different acoustic characteristics and speaker features. For example, some dimensions are used to capture the timbre features of speech, while others focus on depicting the pitch variations of speech, and still others reflect the formant information of speech. This multi-dimensional representation can more comprehensively and meticulously depict voiceprint features. Compared with low-dimensional representations, the 40-dimensional feature vector can accommodate richer information and improve the discriminative power of voiceprint features.
[0074] S33: Obtain the current time and the acquisition time of all voiceprint features in the feature set.
[0075] S34: Calculate the difference between the current time and the acquisition time.
[0076] S35: Calculate the weight of each voiceprint feature in the feature set according to the weighting formula, where the weighting formula is w. t =exp(-λ·Δt), where wt λ represents the weight of the voiceprint features, Δt is the difference between the current time and the acquisition time, and λ is the attenuation factor.
[0077] For steps S31-S35, a relationship between voiceprint feature weights and time differences is constructed by assigning time-decay-based weights to all voiceprint features, including initial and historical voiceprint features. The core logic is that the more timely a voiceprint feature is (i.e., the closer its acquisition time is to the current time), the greater its weight in the feature fusion process; conversely, the weight of a voiceprint feature acquired a long time ago decreases exponentially with increasing time difference. This ensures that recently input voice has a higher weight, facilitating timely reflection of real-time changes in the user's voiceprint. A user's voiceprint changes over time due to changes in health, environmental factors, or device differences. Relatively newer voiceprint features better reflect the user's current voice state and have higher reference value in identity authentication, thus receiving a higher weight to ensure that the fused target voiceprint features better match the user's actual voiceprint situation. By gradually weakening historical voiceprint features, their weights are lowered, preventing the voiceprint database from becoming outdated due to long-term lack of updates to user voiceprints.
[0078] Where Δt is the difference between the current time and the collection time, and in this embodiment, the unit of Δt is days. In this embodiment, λ is the decay factor, used to control the weight decay rate. The default definition of λ is 0.01, but λ can be adjusted according to the business scenario. For example, in a high-frequency trading scenario, λ is set to a value less than 0.01.
[0079] The following are implementation examples of online banking system transfer applications:
[0080] Suppose a user performs a transfer in an online banking system, and their voiceprint database stores three historical voiceprint features, namely voiceprint feature v. a v b v c .
[0081] Voiceprint features v a Data was collected 1 year ago, and the difference between the current time and the collection time is Δt_a = 365 days.
[0082] Voiceprint features v b Data collected 1 month ago, Δt_b = 30 days.
[0083] Voiceprint features v c Data collected today, Δt_c = 0 days.
[0084] If the attenuation factor λ = 0.01 is set, the result can be calculated according to the weighting formula:
[0085] Voiceprint features v aweight w a =exp(-0.01·365)≈0. Due to the long acquisition time, its weight is almost close to 0, and its influence on the target voiceprint features is minimal in feature fusion.
[0086] Voiceprint features v b weight w b =exp(-0.01·30)≈0.7408. This voiceprint feature is relatively new, but still has a certain time difference, so its weight is low.
[0087] Voiceprint features v c weight w c =exp(-0.01·0)≈1, the currently collected voiceprint feature is the most recent, so its weight is the maximum value of 1, playing a dominant role in the feature fusion process. In this way, when performing identity authentication, the system can focus on referring to the latest voiceprint features, making the fused target features closer to the current voiceprint features, avoiding authentication errors due to voiceprint changes or identity theft by criminals using outdated voiceprint features.
[0088] S36: Select voiceprint features with weights greater than the first threshold from the feature set as valid voiceprint features.
[0089] In this embodiment, the first threshold is defined as 0.3, and the weight w t Voiceprint features with a value >0.3 are considered valid voiceprint features, with a weight w. t Voiceprint features with a weight <0.3 will be removed to prevent outdated data from interfering with the voiceprint model. More specifically, if a user's voice is collected once today and once yesterday, generating feature vectors v1 and v2 respectively, both may be retained because their weights are greater than 0.3; however, if a historical voiceprint feature v3 was collected 200 days ago and has a weight w3≈0.14, it will be removed.
[0090] For example, in a securities trading system in the financial sector, a user may have recorded voiceprint features years ago. Over time, these voiceprints may change significantly due to factors such as age and lifestyle changes. These early voiceprint features not only fail to accurately reflect the user's current voice state but may also interfere with the accuracy of voiceprint authentication. By using a first threshold for filtering, such outdated data can be prevented from interfering with the voiceprint model, ensuring that the voiceprint features involved in the fusion are information with high reference value for current authentication, thereby improving the overall quality of voiceprint features. In electronic medical record systems in the medical field, voiceprint features collected during a patient's past illnesses that are no longer representative after recovery can also be excluded through this filtering mechanism, ensuring the accuracy of patient identity authentication and preventing erroneous access to medical record information. It is understood that in other embodiments, the size of the first threshold can be adjusted according to actual needs.
[0091] S37: The target voiceprint feature is obtained by fusing all valid voiceprint features using a weighted average formula; where the weighted average formula is: Where v new The target voiceprint features are n, the number of valid voiceprint features are v. i This is an effective voiceprint feature.
[0092] In step S37, all valid voiceprint features are fused using a weighted average formula to obtain the target voiceprint feature. This fusion method fully considers the weight of each valid voiceprint feature. Voiceprint features with higher weights contribute more to the target voiceprint feature during the weighted average process, while valid voiceprint features with relatively lower weights can also play a supplementary and corrective role. Specifically, although newer historical voiceprint features have slightly lower weights, they can still provide more comprehensive reference information for the target voiceprint feature, making up for the deficiencies of a single voiceprint feature. This allows the fused target voiceprint feature to more accurately reflect the user's voiceprint characteristics and improve the accuracy of identity authentication.
[0093] The following is an implementation example of target voiceprint feature generation when a user uses a financial application platform or a medical financial platform, and voiceprint drift caused by a cold:
[0094] Day 1 (early stage of a cold, sudden change in voiceprint):
[0095] Acquire new input speech, extract the voiceprint feature vector v1, calculate its weight as w1 = exp(-0.01·0) ≈ 1 > 0.3, retain this voiceprint feature vector v1 as the effective voiceprint feature, and obtain the target voiceprint feature. Therefore, the voiceprint database was updated to the voiceprint features v1 of a person with a cold.
[0096] Day 7 (Cold recovery period, voiceprint gradually changing):
[0097] As the user recovers from a cold and their voiceprint gradually returns to normal, the input speech is collected, and the voiceprint feature vector v2 is extracted. Its weight is calculated as w2 = exp(-0.01·0) ≈ 1 > 0.3. At this point, w′1 = exp(-0.01·7) ≈ 0.932 > 0.3. Voiceprint feature vectors v1 and v2 are retained as valid voiceprint features, resulting in the target voiceprint feature v. new_2 = [v2+0.932v1] / [1+0.932]≈0.5176v2+0.4824v1, the voiceprint library is transitioning to the recovery period voiceprint.
[0098] Day 122 (Cold fully recovered, voiceprint normal):
[0099] The user's voiceprint is fully recovered and returns to normal. Input speech is collected, and voiceprint feature vector v3 is extracted. Its weight is calculated as w3 = exp(-0.01·0) ≈ 1 > 0.3. Simultaneously, the weights of voiceprint feature vectors v1 and v2 are w″1 = exp(-0.01·122) ≈ 0.296 < 0.3 and w′2 = exp(-0.01·115) ≈ 0.316 > 0.3, respectively. Voiceprint feature vectors v2 and v3 are retained as valid voiceprint features, resulting in the target voiceprint feature v. new_3 = [v3+0.316v2] / [1+0.316]≈0.76v3+0.24v2, triggering the voiceprint model to remove historical voiceprint features, ensuring the timeliness of the voiceprint database, so that the voiceprint features of the voiceprint database are updated to be dominated by normal voiceprints, and can ensure that even if you catch a cold again recently, you can still pass the verification, avoiding verification failure due to temporary voiceprint changes. The voiceprint model returns to the normal state with the voiceprint changes, without the need for manual intervention.
[0100] Specifically, after fusing the extracted voiceprint features with the historical voiceprint features in the voiceprint database to generate the target voiceprint features, i.e. after step S30, the method further includes: using an incremental learning model to learn the target voiceprint features and fine-tuning the historical voiceprint features to update the voiceprint database.
[0101] Understandably, in this implementation, the voiceprint model is a model for voiceprint authentication built based on voiceprint features in the voiceprint database; the voiceprint database is a database that stores users' historical voiceprint features, providing data support for the voiceprint model; the incremental learning model is an algorithm that can continuously learn and update the model based on target voiceprint features. Traditional static voiceprint model updates often require a complete update of the voiceprint database, which is not only time-consuming and laborious, but may also be subject to many limitations in practical applications. The incremental learning model, on the other hand, can selectively update the voiceprint model locally for newly emerging target voiceprint features, and fine-tune historical voiceprint features, such as removing the historical voiceprint features collected the oldest time. Specifically, when a target voiceprint feature needs to be included in the voiceprint model, the incremental learning model will first analyze the target voiceprint feature, compare it with the existing voiceprint model, and then, based on the differences in the target voiceprint feature, adjust and optimize only the parts of the model related to these features. This approach allows for the timely integration of new voiceprint information into the model, improving the accuracy and adaptability of voiceprint recognition. It also avoids large-scale updates to the entire voiceprint database, significantly reducing computational and storage requirements and improving system efficiency. This enables the voiceprint recognition system to adapt more efficiently and flexibly to constantly changing voiceprint data. Fine-tuning historical voiceprint features and updating the voiceprint database ensures that the data remains highly valuable. For example, as a user's voiceprint gradually recovers, such as after a complete recovery from a cold, the voiceprint features in the database will be updated to be dominated by normal voiceprints. This allows the voiceprint model to return to a normal state following voiceprint changes, preventing authentication errors due to outdated data and helping to ensure the accuracy of user authentication on financial and medical application platforms.
[0102] In this embodiment, a lightweight compressed Transformer is used as the incremental learning model. This model structure is simplified to a 3-layer encoder, with the number of heads in each layer reduced from 8 to 4, thus reducing computational complexity. The lightweight compressed Transformer is an optimized and improved model structure of the traditional Transformer model. The Transformer model, a deep learning model proposed by Google, uses self-attention as its core technology, revolutionizing the field of natural language processing. The Transformer model consists of an encoder and a decoder, each containing multiple stacked layers. The lightweight compressed Transformer reduces the model's storage space and computational load while maintaining performance through adjustments to the model architecture and parameter compression, making it more suitable for operation on resource-constrained devices or in scenarios.
[0103] The loss function of the incremental learning model in this embodiment is: Where α = 0.7, β = 0.3, θold For historical model parameters, α·L CE (D new ,θ) is the target voiceprint feature classification loss, As a parameter stability constraint, the setting of this loss function can avoid model oscillations caused by incremental updates.
[0104] S40: Perform identity authentication matching on the target voiceprint features and generate preliminary authentication results.
[0105] For step S40, a similarity calculation and dynamic threshold determination method can be used to perform identity authentication matching of the target voiceprint features. The similarity calculation and dynamic threshold determination method are existing technologies and will not be elaborated upon here. In specific implementation, if a user conducts a voice-controlled transfer transaction through mobile banking or online banking, during the transfer process, the user needs to state information such as the transaction recipient, transaction type, and transaction amount. The system collects the user's voiceprint data and extracts the target voiceprint features. The similarity calculation and dynamic threshold determination method is then used to match the target voiceprint features with the user's pre-stored voiceprint samples for identity authentication. If the similarity is higher than the dynamic threshold, the initial authentication result is successful, and the user can continue to complete the account opening process. If the similarity is lower than the threshold, the authentication result is unsuccessful, and the bank will require the user to perform voice control again, re-collect voiceprints, or use other identity verification methods. For example, when doctors are making rounds or conducting telemedicine consultations, they need to access patients' electronic medical records. Through a voice command system, the doctor states their identity information and access request. The system collects the doctor's voiceprint and extracts the target voiceprint features, matching them with the doctor's voiceprint database stored in the hospital's information system for identity authentication. Using similarity calculation and dynamic threshold determination, if the authentication is successful, the doctor can obtain access to the corresponding patient's electronic medical records. If it fails, the system will refuse access and prompt the doctor that the identity verification failed, requiring the doctor to re-verify the voiceprint or use other identity recognition methods to log in to the system.
[0106] S50: Perform security verification on the input voice and generate a security verification result.
[0107] This embodiment employs a dual mechanism of identity authentication matching and voice security verification to comprehensively defend against attacks on voiceprint authentication, build a security defense for financial transactions and medical data management, maintain normal transaction order and medical environment, and enhance users' and institutions' trust in voiceprint authentication technology.
[0108] Specifically, security verification includes: semantic consistency detection, synthesized speech detection, and recording attack detection. Semantic consistency is used to detect whether the user's input voice matches the actual operation information, preventing erroneous transaction operations due to unclear pronunciation or being misled, ensuring the accuracy and security of authentication operations, and maintaining normal transaction order. Synthetic speech detection and recording attack detection are used to determine whether the input voice is AI-synthesized or recorded, resisting attacks that forge voiceprints and maliciously use historical voiceprint information for identity authentication, protecting the user's identity information and asset security, preventing attackers from gaining access and operation rights to financial transaction or medical data management systems, and enhancing the security and reliability of the voiceprint authentication system.
[0109] In some embodiments of this application, as shown in FIG5, a specific security verification scheme is provided. In S50, the input voice is verified for security and a security verification result is generated, which specifically includes the following steps S51-S55.
[0110] S51: Parse the semantic content of the input speech and detect the correlation between the semantic content and the actual transaction operation.
[0111] When a user's pronunciation is unclear, the speech recognition system may make a recognition error. Step S51 analyzes the semantic content. If the user's unclear pronunciation leads to semantic ambiguity and has a low correlation with the actual transaction operation judged by the system, the user needs to input speech for secondary verification. If the correlation is high, the input speech and semantic content can be judged to be valid, and the corresponding operation can be responded to.
[0112] In some embodiments of this application, as shown in FIG6, a specific semantic consistency detection scheme is provided. In S51, that is, parsing the semantic content of the input speech and detecting the correlation between the semantic content and the actual transaction operation, the specific steps are as follows: S511-S515.
[0113] S511: Uses a pre-trained language model with bidirectional encoding to parse the semantic content of the input speech.
[0114] In this embodiment, a pre-trained BERT model is used to parse user speech content. BERT (Bidirectional Encoder Representations from Transformers) is a bidirectional pre-trained language model proposed by Google in 2018. Based on the Transformer architecture, it aims to capture deep semantic relationships in text through large-scale unsupervised learning. Its core innovation lies in bidirectional context modeling, which can simultaneously analyze the contextual information of each word in a sentence. Different BERT models can be pre-trained depending on the application domain. For example, in financial applications, a pre-trained BERT model for the financial domain is used. The financial domain contains a large number of professional terms, specific concepts, and complex semantic relationships, such as "hedging" and "asset securitization." A pre-trained BERT model for the financial domain can learn these professional terms and text patterns, better understand the semantics of financial texts, and accurately process and analyze financial-related information, such as contract terms, financial reports, and market commentary, thereby providing more accurate support for various financial operations. For example, in medical applications, a pre-trained BERT model for the medical field can be used. The medical field contains a large number of professional terms, abbreviations, and specific medical concepts, such as "myocardial infarction" and "CT scan." A pre-trained BERT model for the medical field can learn these professional terms and text patterns, accurately understand the semantics of medical texts such as electronic medical records, medical literature, and examination reports. Medical data contains a large amount of unstructured text, such as doctors' diagnostic records and patients' symptom descriptions. The BERT model for the medical field can quickly and accurately extract key information from these texts, such as disease symptoms, diagnosis results, and treatment plans, which helps to improve the accuracy of judging user intent.
[0115] S512: Extract the transaction intent and key entities from the voice content, and concatenate the transaction intent and key entities into target operation information.
[0116] Understandably, in the financial application field, the output transaction intent I∈{transfer, query, account opening, ...} and key entities (such as amount, payee) can be combined to form the target operation information "transfer 3000 yuan to Zhang San". In the medical application field, the output transaction intent I∈{query, appointment, ...} and key entities (such as time, department) can be combined to form the target operation information "appoint an appointment for the internal medicine outpatient clinic at 9 am tomorrow".
[0117] S513: Encode the target operation information and output the embedding vector of the target operation information.
[0118] Step S513 can extract and splice the target operation information obtained from the speech content, such as "transfer 3,000 yuan to Zhang San" and "make an appointment for the internal medicine clinic at 9 am tomorrow", and convert it into a numerical form that the computer can process, namely, an embedded vector. This realizes the structuring and digitization of the target operation information, which makes it easy to compare and analyze it with the actual operation information in the future, laying the foundation for calculating the correlation between the two.
[0119] S514: An embedding vector for obtaining actual user operation information.
[0120] S515: Calculate the correlation between the target operation information and the actual operation information according to the semantic consistency scoring formula, where the semantic consistency scoring formula is S sem =cos(E speech E operation ), where S sem To establish the correlation between target operational information and actual operational information, E speech and E operation These are the embedding vectors of the target operation information and the actual operation information, respectively.
[0121] For steps S514-S515, the correlation between the target operation information and the actual operation information is calculated according to the semantic consistency scoring formula. By calculating the relationship between the two embedding vectors, such as the cosine similarity in this embodiment, a quantitative correlation score is obtained. This score can intuitively reflect the degree of semantic matching between the target operation information and the actual operation information, thereby judging the degree of correlation between the semantic content of the user's input voice and the actual transaction operation.
[0122] S52: Perform synthesized speech detection and recording attack identification on the input speech to obtain the attack probability of the input speech.
[0123] This embodiment uses a Generative Adversarial Network (GAN) to construct a detection model for synthetic speech detection. This model consists of a generator G and a discriminator D. The generator G aims to generate synthetic speech that is as realistic as possible to deceive the discriminator D; while the discriminator D aims to accurately distinguish between real speech and synthetic / recorded attack speech. During training, the discriminator D receives two types of input: real speech samples, i.e., normal user speech collected from real-world application scenarios or standard speech datasets; and synthetic speech samples generated by the generator G, i.e., fake speech generated by the generator G based on patterns learned from training data. This embodiment optimizes the discriminator D parameters by minimizing the discriminator D's loss function, where the loss function of the discriminator D is:
[0124] Where E represents the expectation, x is the real speech sample, and p realLet z be the data distribution of real speech, and p be random noise. z Let G(z) be the noise distribution and G(z) be the synthesized speech generated by the generator. The loss function aims to make the discriminator D classify real speech as "real" as possible, i.e., maximizing logD(x), and classify synthesized speech as "fake", i.e., maximizing log(1-D(G(z))). After training, the input speech is fed into the trained discriminator D for detection. The discriminator D outputs a probability value, representing the probability that the input speech is synthesized. This probability value ranges from 0 to 1; a higher value indicates that the input speech is more likely to be attack speech, and vice versa. In addition, this embodiment uses publicly available synthesized speech datasets, such as ASVSpoof2019, for model training. The ASVSpoof 2019 dataset contains a large number of real speech samples as well as synthesized speech samples generated by different techniques and recorded attack samples, providing data support for the model to learn to distinguish between real and fake speech.
[0125] This embodiment calculates the spectral differences between adjacent speech segments by analyzing the temporal continuity of the speech signal:
[0126] Where S t Let t be the spectrogram of the t-th frame. The Δspec of natural speech is usually small, typically less than 0.05. However, the Δspec of recording attacks increases significantly due to sudden changes in background noise or splicing. Therefore, the difference in the spectrogram of speech segments in adjacent frames can be used to determine whether the input speech is a recording.
[0127] In this embodiment, input speech with a synthesis probability greater than or equal to 0.9 is defined as synthesized speech, and input speech with a spectral difference between adjacent frame speech segments greater than or equal to 0.05 is defined as recorded speech.
[0128] S53: Compare the correlation with the second threshold and the attack probability with the third threshold.
[0129] In this embodiment, the second threshold is defined as 0.6. If it is less than 0.6, the verification condition for semantic consistency detection is not met. A third threshold can be defined as 0.5. If the synthesis probability is greater than or equal to 0.9 and the spectral difference between adjacent frame speech segments is greater than or equal to 0.05, the attack probability can be defined as exceeding 0.5; otherwise, the attack probability can be defined as less than 0.5.
[0130] S54: If the correlation of the input voice is less than the second threshold or the attack probability is greater than the third threshold, then output a security verification failure signal.
[0131] S55: If the correlation of the input voice is greater than the second threshold and the attack probability is less than the third threshold, then output a security verification pass signal.
[0132] For steps S54-S55, a second threshold and a third threshold are set to quantitatively assess the semantic relevance of the input speech and the presence of a speech attack. When the relevance of the input speech is less than the second threshold, it indicates a low degree of matching between the semantic content of the user's input speech and the actual transaction operation, potentially indicating misunderstanding, misoperation, or malicious inducement. When the attack probability is greater than the third threshold, it indicates a high probability that the input speech is synthesized speech or a recorded attack, posing a security risk. In this case, a security verification failure signal is output, rejecting the user's operation request to ensure the security of the system and the user. Conversely, if the relevance of the input speech is greater than the second threshold and the attack probability is less than the third threshold, it indicates that the speech is semantically consistent with the operation and there is no obvious attack risk. In this case, a security verification pass signal is output, providing security for subsequent operations.
[0133] S60: Output the identity authentication result based on the security verification result and the preliminary authentication result.
[0134] Specifically, for step S60, if the initial authentication fails or the security verification fails, an identity authentication failure result is output and a secondary authentication signal is sent; if the initial authentication and the security verification pass, an identity authentication pass result is output.
[0135] As can be seen, the above solution automatically completes voiceprint authentication and updates without the user's awareness by collecting voice in real time, significantly improving the user experience without requiring the user to actively update their voiceprint. By fusing the extracted voiceprint features with historical features, dynamic updates of voiceprint features are achieved, allowing the voiceprint authentication method to automatically adapt to voiceprint changes caused by factors such as age and health. Compared with traditional static voiceprint databases, this significantly improves the matching rate, reduces the number of re-entry attempts, and greatly optimizes the user experience. At the same time, it ensures the operational efficiency of financial and medical services, and the dynamic update mechanism enhances the security of the voiceprint database, reduces the risk of attacks due to outdated voiceprint information, and effectively protects user funds and privacy. By adopting a dual mechanism of identity authentication matching and voice security verification, it comprehensively resists attacks on voiceprint authentication, builds a security defense for financial transactions and medical data management, maintains normal transaction order and the medical environment, and enhances the trust of users and institutions in voiceprint authentication technology.
[0136] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0137] In one embodiment, this application provides a dynamic voiceprint authentication device, which corresponds one-to-one with the dynamic voiceprint authentication method in the above embodiments. As shown in FIG7, the dynamic voiceprint authentication device includes a data acquisition module 101, an extraction module 102, a fusion module 103, a matching module 104, a verification module 105, and an authentication module 106. Detailed descriptions of each functional module are as follows:
[0138] The acquisition module 101 is used to acquire input voice in real time and trigger an authentication command when input voice is acquired;
[0139] Extraction module 102 is used to preprocess the input speech and extract voiceprint features;
[0140] The fusion module 103 is used to fuse the extracted voiceprint features with historical voiceprint features to generate target voiceprint features;
[0141] The matching module 104 is used to perform identity authentication matching on the target voiceprint features and generate preliminary authentication results;
[0142] The verification module 105 is used to perform security verification on the input voice and generate a security verification result;
[0143] The authentication module 106 is used to output the identity authentication result based on the security verification result and the preliminary authentication result.
[0144] In one embodiment, the extraction module 102 is specifically used for:
[0145] The input speech is subjected to spectral subtraction noise reduction and frame segmentation to obtain several speech segments;
[0146] Mel frequency cepstral coefficients and linear prediction coding features are extracted sequentially from each frame of speech to obtain several initial speaker features.
[0147] In one embodiment, the fusion module 103 is specifically used for:
[0148] Retrieve historical voiceprint features from the corresponding user's voiceprint database;
[0149] The initial voiceprint features and historical voiceprint features are combined to form a feature set;
[0150] Obtain the current time and the acquisition time of all voiceprint features in the feature set;
[0151] Calculate the difference between the current time and the acquisition time;
[0152] The weight of each voiceprint feature in the feature set is calculated according to the weighting formula, where the weighting formula is w. t =exp(-λ·Δt), where w tλ represents the weight of the voiceprint features, Δt is the difference between the current time and the acquisition time, and λ is the attenuation factor.
[0153] Voiceprint features with weights greater than the first threshold in the feature set are selected as valid voiceprint features.
[0154] The target voiceprint feature is obtained by fusing all valid voiceprint features using a weighted average formula; where the weighted average formula is: Where v new The target voiceprint features are n, the number of valid voiceprint features are v. i This is an effective voiceprint feature.
[0155] In one embodiment, the authentication module 106 is specifically used for:
[0156] Parse the semantic content of the input speech and detect the correlation between the semantic content and the actual transaction operation;
[0157] Synthetic speech detection and recording attack identification are performed on the input speech to obtain the attack probability of the input speech;
[0158] The correlation is compared with the second threshold, and the attack probability is compared with the third threshold.
[0159] If the relevance of the input speech is less than the second threshold or the attack probability is greater than the third threshold, a security verification failure signal will be output.
[0160] If the relevance of the input speech is greater than the second threshold and the attack probability is less than the third threshold, then a security verification pass signal is output.
[0161] Specific limitations regarding the dynamic voiceprint authentication device can be found in the limitations of the dynamic voiceprint authentication method above, and will not be repeated here. Each module in the aforementioned dynamic voiceprint authentication device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0162] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram is shown in Figure 8. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of the computer device is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a dynamic voiceprint authentication method server.
[0163] In one embodiment, a computer device is provided, which can be a client, and its internal structure diagram is shown in Figure 9. The computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a dynamic voiceprint authentication method on the client side.
[0164] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:
[0165] Real-time acquisition of input voice and triggering an authentication command when input voice is acquired;
[0166] Preprocess the input speech and extract speaker features;
[0167] The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features;
[0168] The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated.
[0169] Perform security verification on the input voice and generate the security verification result;
[0170] The identity authentication result is output based on the security verification result and the preliminary authentication result.
[0171] In one embodiment, a non-volatile computer-readable storage medium is provided, on which a computer program is stored, the computer program performing the following steps when executed by a processor:
[0172] Real-time acquisition of input voice and triggering an authentication command when input voice is acquired;
[0173] Preprocess the input speech and extract speaker features;
[0174] The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features;
[0175] The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated.
[0176] Perform security verification on the input voice and generate the security verification result;
[0177] The identity authentication result is output based on the security verification result and the preliminary authentication result.
[0178] It should be noted that the functions or steps that can be implemented by the non-volatile computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0180] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0181] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A dynamic voiceprint authentication method, comprising: Real-time acquisition of input voice and triggering an authentication command when input voice is acquired; Preprocess the input speech and extract speaker features; The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features; The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated. Perform security verification on the input voice and generate the security verification result; The identity authentication result is output based on the security verification result and the preliminary authentication result.
2. The dynamic voiceprint authentication method according to claim 1, wherein, The preprocessing of the input speech and extraction of voiceprint features includes: The input speech is subjected to spectral subtraction noise reduction and frame segmentation to obtain several speech segments; Mel frequency cepstral coefficients and linear prediction coding features are extracted sequentially from each frame of speech to obtain several initial speaker features.
3. The dynamic voiceprint authentication method according to claim 2, wherein, The process of fusing the extracted voiceprint features with historical voiceprint features to generate target voiceprint features includes: Retrieve historical voiceprint features from the corresponding user's voiceprint database; The initial voiceprint features and historical voiceprint features are combined to form a feature set; Obtain the current time and the acquisition time of all voiceprint features in the feature set; Calculate the difference between the current time and the acquisition time; The weight of each voiceprint feature in the feature set is calculated according to the weighting formula, where the weighting formula is w. tt =exp(-λ·Δt), where w t λ represents the weight of the voiceprint features, Δt is the difference between the current time and the acquisition time, and λ is the attenuation factor.
4. The dynamic voiceprint authentication method according to claim 3, wherein, The step of fusing the extracted voiceprint features with historical voiceprint features to generate the target voiceprint features also includes: Voiceprint features with weights greater than the first threshold in the feature set are selected as valid voiceprint features. The target voiceprint feature is obtained by fusing all valid voiceprint features using a weighted average formula; where the weighted average formula is: Where v new The target voiceprint features are n, the number of valid voiceprint features are v. i This is an effective voiceprint feature.
5. The dynamic voiceprint authentication method according to claim 1, wherein, After fusing the extracted voiceprint features with historical voiceprint features in the voiceprint database to generate the target voiceprint features, the process further includes: An incremental learning model is used to learn the target voiceprint features, and the historical voiceprint features are fine-tuned to update the voiceprint database.
6. The dynamic voiceprint authentication method according to claim 1, wherein, The step of performing security verification on the input voice and generating a security verification result includes: Parse the semantic content of the input speech and detect the correlation between the semantic content and the actual transaction operation; Synthetic speech detection and recording attack identification are performed on the input speech to obtain the attack probability of the input speech; The correlation is compared with the second threshold, and the attack probability is compared with the third threshold. If the relevance of the input speech is less than the second threshold or the attack probability is greater than the third threshold, a security verification failure signal will be output. If the relevance of the input speech is greater than the second threshold and the attack probability is less than the third threshold, then a security verification pass signal is output.
7. The dynamic voiceprint authentication method according to claim 6, wherein, The process of performing security verification on the target voiceprint features and outputting the identity authentication result based on the verification result includes: A pre-trained language model with bidirectional encoding is used to parse the semantic content of the input speech; Extract transaction intent and key entities from the voice content, and then concatenate the transaction intent and key entities into target operation information; The target operation information is encoded, and the embedding vector of the target operation information is output. Obtain the embedding vector of the user's actual operation information; The correlation between target operation information and actual operation information is calculated based on the semantic consistency scoring formula, where the semantic consistency scoring formula is S. sem =cos(E speech E operation ), where S sem To establish the correlation between target operational information and actual operational information, E speech and E operation These are the embedding vectors of the target operation information and the actual operation information, respectively.
8. A dynamic voiceprint authentication device, comprising: The acquisition module is used to acquire input voice in real time and trigger an authentication command when input voice is acquired; The extraction module is used to preprocess the input speech and extract speaker features; The fusion module is used to fuse the extracted voiceprint features with historical voiceprint features to generate the target voiceprint features; The matching module is used to perform identity authentication matching on the target voiceprint features and generate preliminary authentication results; The verification module is used to perform security verification on the input voice and generate the security verification result; The authentication module is used to output the identity authentication result based on the security verification result and the preliminary authentication result.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the computer program, it performs the following steps: Real-time acquisition of input voice and triggering an authentication command when input voice is acquired; Preprocess the input speech and extract speaker features; The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features; The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated. Perform security verification on the input voice and generate the security verification result; The identity authentication result is output based on the security verification result and the preliminary authentication result.
10. The computer device according to claim 9, wherein, The preprocessing of the input speech and extraction of voiceprint features includes: The input speech is subjected to spectral subtraction noise reduction and frame segmentation to obtain several speech segments; Mel frequency cepstral coefficients and linear prediction coding features are extracted sequentially from each frame of speech to obtain several initial speaker features.
11. The computer device according to claim 10, wherein, The process of fusing the extracted voiceprint features with historical voiceprint features to generate target voiceprint features includes: Retrieve historical voiceprint features from the corresponding user's voiceprint database; The initial voiceprint features and historical voiceprint features are combined to form a feature set; Obtain the current time and the acquisition time of all voiceprint features in the feature set; Calculate the difference between the current time and the acquisition time; The weight of each voiceprint feature in the feature set is calculated according to the weighting formula, where the weighting formula is w. t =exp(-λ·Δt), where w t λ represents the weight of the voiceprint features, Δt is the difference between the current time and the acquisition time, and λ is the attenuation factor.
12. The computer device according to claim 11, wherein, The step of fusing the extracted voiceprint features with historical voiceprint features to generate the target voiceprint features also includes: Voiceprint features with weights greater than the first threshold in the feature set are selected as valid voiceprint features. The target voiceprint feature is obtained by fusing all valid voiceprint features using a weighted average formula; where the weighted average formula is: Where v new The target voiceprint features are n, the number of valid voiceprint features are v. i This is an effective voiceprint feature.
13. The computer device according to claim 9, wherein, After fusing the extracted voiceprint features with historical voiceprint features in the voiceprint database to generate the target voiceprint features, the process further includes: An incremental learning model is used to learn the target voiceprint features, and the historical voiceprint features are fine-tuned to update the voiceprint database.
14. The computer device according to claim 9, wherein, The step of performing security verification on the input voice and generating a security verification result includes: Parse the semantic content of the input speech and detect the correlation between the semantic content and the actual transaction operation; Synthetic speech detection and recording attack identification are performed on the input speech to obtain the attack probability of the input speech; The correlation is compared with the second threshold, and the attack probability is compared with the third threshold. If the relevance of the input speech is less than the second threshold or the attack probability is greater than the third threshold, a security verification failure signal will be output. If the relevance of the input speech is greater than the second threshold and the attack probability is less than the third threshold, then a security verification pass signal is output.
15. The computer device according to claim 14, wherein, The process of performing security verification on the target voiceprint features and outputting the identity authentication result based on the verification result includes: A pre-trained language model with bidirectional encoding is used to parse the semantic content of the input speech; Extract transaction intent and key entities from the voice content, and then concatenate the transaction intent and key entities into target operation information; The target operation information is encoded, and the embedding vector of the target operation information is output. Obtain the embedding vector of the user's actual operation information; The correlation between target operation information and actual operation information is calculated based on the semantic consistency scoring formula, where the semantic consistency scoring formula is S. sem =cos(E speech E operation ), where S sem To establish the correlation between target operational information and actual operational information, E speech and E operation These are the embedding vectors of the target operation information and the actual operation information, respectively.
16. A non-volatile computer-readable storage medium storing a computer program, wherein, When the computer program is executed by the processor, it performs the following steps: Real-time acquisition of input voice and triggering an authentication command when input voice is acquired; Preprocess the input speech and extract speaker features; The extracted voiceprint features are fused with historical voiceprint features to generate the target voiceprint features; The target voiceprint features are matched for identity authentication, and a preliminary authentication result is generated. Perform security verification on the input voice and generate the security verification result; The identity authentication result is output based on the security verification result and the preliminary authentication result.
17. The non-volatile computer-readable storage medium according to claim 16, wherein, The preprocessing of the input speech and extraction of voiceprint features includes: The input speech is subjected to spectral subtraction noise reduction and frame segmentation to obtain several speech segments; Mel frequency cepstral coefficients and linear prediction coding features are extracted sequentially from each frame of speech to obtain several initial speaker features.
18. The non-volatile computer-readable storage medium according to claim 17, wherein, The process of fusing the extracted voiceprint features with historical voiceprint features to generate target voiceprint features includes: Retrieve historical voiceprint features from the corresponding user's voiceprint database; The initial voiceprint features and historical voiceprint features are combined to form a feature set; Obtain the current time and the acquisition time of all voiceprint features in the feature set; Calculate the difference between the current time and the acquisition time; The weight of each voiceprint feature in the feature set is calculated according to the weighting formula, where the weighting formula is w. t =exp(-λ·Δt), where w t λ represents the weight of the voiceprint features, Δt is the difference between the current time and the acquisition time, and λ is the attenuation factor.
19. The non-volatile computer-readable storage medium according to claim 18, wherein, The step of fusing the extracted voiceprint features with historical voiceprint features to generate the target voiceprint features also includes: Voiceprint features with weights greater than the first threshold in the feature set are selected as valid voiceprint features. The target voiceprint feature is obtained by fusing all valid voiceprint features using a weighted average formula; where the weighted average formula is: Where v new The target voiceprint features are n, the number of valid voiceprint features are v. i This is an effective voiceprint feature.
20. The non-volatile computer-readable storage medium according to claim 16, wherein, After fusing the extracted voiceprint features with historical voiceprint features in the voiceprint database to generate the target voiceprint features, the process further includes: An incremental learning model is used to learn the target voiceprint features, and the historical voiceprint features are fine-tuned to update the voiceprint database.
